Details

Autor(en) / Beteiligte

• Ravishankar, Anand
• Natarajan, Santhi
• Bharathi Malakreddy, A.
• Jung, Matthias
• Orailoglu, Alex
• Reichenbach, Marc

Titel

SparseMAX: Accelerating Quantum Neural Networks on GPU Clusters Using Sparse-Matrix Kernels

Ist Teil von

• Embedded Computer Systems: Architectures, Modeling, and Simulation, p.418-431

Ort / Verlag

Cham: Springer International Publishing

Quelle

Alma/SFX Local Collection

Beschreibungen/Notizen

• The growing popularity of Applied Quantum Mechanics and Artificial Intelligence drives the need for integrating the two fields. Quantum Neural Networks (QNNs) incorporate quantum aspects into classical deep learning networks which are capable of performing universal quantum computations. The dense representation of QNNs presents great challenges in terms of computational cost and noise susceptibility. In this paper, we present SparseMAX, a novel Sparse Quantum Neural Network (SQNN) that is robust to noise and interference for large volumes of network parameters. We also introduce Quantron (ψ\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\psi $$\end{document}), a generalized version of perceptron, which acts on qubits and performs the necessary quantum operations. Based on these insights, we develop 2 GPU kernels. The first kernel estimates the network architecture through a quantum training algorithm. The second kernel accelerates a sparsified version of the network matrices on a GPU cluster. We validate our kernel performance and training algorithm and present the results in terms of inference time, GPU efficiency and scalability. On an average, SparseMAX utilizes 54.83% of our GPU cluster’s compute resources, while offering a 41.51×\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times $$\end{document} speedup in terms of serial inference timing measurements for network layer range [120, 1920] and neurons per layer range [1024,4096]